Modern software engineering faces challenges in task assignment due to over-reliance on static identifiers, which ignore contextual data like documentation, commit logs, and developer profiles. This limits adaptability in resource allocation and bug triage. We propose SODAOpt, a transformer-based framework that integrates textual and socio-demographic embeddings to enhance task matching. Our approach employs adaptive fusion layers to unify heterogeneous data modalities and a composite loss function balancing contrastive learning with assignment optimization. Evaluations on real-world datasets demonstrate SODAOpt’s superiority in precision and cross-project diversity, validated by tailored metrics (SP@K, CPR, C-NDCG@K).